On Fri, Dec 6, 2024 at 11:30 PM Andrii Nakryiko
<andrii.nakryiko@xxxxxxxxx> wrote:
> On Fri, Dec 6, 2024 at 2:25 PM Jann Horn <jannh@xxxxxxxxxx> wrote:
> >
> > On Fri, Dec 6, 2024 at 11:15 PM Andrii Nakryiko
> > <andrii.nakryiko@xxxxxxxxx> wrote:
> > > On Fri, Dec 6, 2024 at 12:45 PM Jann Horn <jannh@xxxxxxxxxx> wrote:
> > > >
> > > > Currently, the pointer stored in call->prog_array is loaded in
> > > > __uprobe_perf_func(), with no RCU annotation and no RCU protection, so the
> > > > loaded pointer can immediately be dangling. Later,
> > > > bpf_prog_run_array_uprobe() starts a RCU-trace read-side critical section,
> > > > but this is too late. It then uses rcu_dereference_check(), but this use of
> > > > rcu_dereference_check() does not actually dereference anything.
> > > >
> > > > It looks like the intention was to pass a pointer to the member
> > > > call->prog_array into bpf_prog_run_array_uprobe() and actually dereference
> > > > the pointer in there. Fix the issue by actually doing that.
> > > >
> > > > Fixes: 8c7dcb84e3b7 ("bpf: implement sleepable uprobes by chaining gps")
> > > > Cc: stable@xxxxxxxxxxxxxxx
> > > > Signed-off-by: Jann Horn <jannh@xxxxxxxxxx>
> > > > ---
> > > > To reproduce, in include/linux/bpf.h, patch in a mdelay(10000) directly
> > > > before the might_fault() in bpf_prog_run_array_uprobe() and add an
> > > > include of linux/delay.h.
> > > >
> > > > Build this userspace program:
> > > >
> > > > ```
> > > > $ cat dummy.c
> > > > #include <stdio.h>
> > > > int main(void) {
> > > > printf("hello world\n");
> > > > }
> > > > $ gcc -o dummy dummy.c
> > > > ```
> > > >
> > > > Then build this BPF program and load it (change the path to point to
> > > > the "dummy" binary you built):
> > > >
> > > > ```
> > > > $ cat bpf-uprobe-kern.c
> > > > #include <linux/bpf.h>
> > > > #include <bpf/bpf_helpers.h>
> > > > #include <bpf/bpf_tracing.h>
> > > > char _license[] SEC("license") = "GPL";
> > > >
> > > > SEC("uprobe//home/user/bpf-uprobe-uaf/dummy:main")
> > > > int BPF_UPROBE(main_uprobe) {
> > > > bpf_printk("main uprobe triggered\n");
> > > > return 0;
> > > > }
> > > > $ clang -O2 -g -target bpf -c -o bpf-uprobe-kern.o bpf-uprobe-kern.c
> > > > $ sudo bpftool prog loadall bpf-uprobe-kern.o uprobe-test autoattach
> > > > ```
> > > >
> > > > Then run ./dummy in one terminal, and after launching it, run
> > > > `sudo umount uprobe-test` in another terminal. Once the 10-second
> > > > mdelay() is over, a use-after-free should occur, which may or may
> > > > not crash your kernel at the `prog->sleepable` check in
> > > > bpf_prog_run_array_uprobe() depending on your luck.
> > > > ---
> > > > Changes in v2:
> > > > - remove diff chunk in patch notes that confuses git
> > > > - Link to v1: https://lore.kernel.org/r/20241206-bpf-fix-uprobe-uaf-v1-1-6869c8a17258@xxxxxxxxxx
> > > > ---
> > > > include/linux/bpf.h | 4 ++--
> > > > kernel/trace/trace_uprobe.c | 2 +-
> > > > 2 files changed, 3 insertions(+), 3 deletions(-)
> > > >
> > >
> > > Looking at how similar in spirit bpf_prog_run_array() is meant to be
> > > used, it seems like it is the caller's responsibility to
> > > RCU-dereference array and keep RCU critical section before calling
> > > into bpf_prog_run_array(). So I wonder if it's best to do this instead
> > > (Gmail will butcher the diff, but it's about the idea):
> >
> > Yeah, that's the other option I was considering. That would be more
> > consistent with the existing bpf_prog_run_array(), but has the
> > downside of unnecessarily pushing responsibility up to the caller...
> > I'm fine with either.
>
> there is really just one caller ("legacy" singular uprobe handler), so
> I think this should be fine. Unless someone objects I'd keep it
> consistent with other "prog_array_run" helpers
Ack, I will make it consistent.
> > > diff --git a/include/linux/bpf.h b/include/linux/bpf.h
> > > index eaee2a819f4c..4b8a9edd3727 100644
> > > --- a/include/linux/bpf.h
> > > +++ b/include/linux/bpf.h
> > > @@ -2193,26 +2193,25 @@ bpf_prog_run_array(const struct bpf_prog_array *array,
> > > * rcu-protected dynamically sized maps.
> > > */
> > > static __always_inline u32
> > > -bpf_prog_run_array_uprobe(const struct bpf_prog_array __rcu *array_rcu,
> > > +bpf_prog_run_array_uprobe(const struct bpf_prog_array *array,
> > > const void *ctx, bpf_prog_run_fn run_prog)
> > > {
> > > const struct bpf_prog_array_item *item;
> > > const struct bpf_prog *prog;
> > > - const struct bpf_prog_array *array;
> > > struct bpf_run_ctx *old_run_ctx;
> > > struct bpf_trace_run_ctx run_ctx;
> > > u32 ret = 1;
> > >
> > > might_fault();
> > > + RCU_LOCKDEP_WARN(!rcu_read_lock_trace_held(), "no rcu lock held");
> > > +
> > > + if (unlikely(!array))
> > > + goto out;
> > >
> > > - rcu_read_lock_trace();
> > > migrate_disable();
> > >
> > > run_ctx.is_uprobe = true;
> > >
> > > - array = rcu_dereference_check(array_rcu, rcu_read_lock_trace_held());
> > > - if (unlikely(!array))
> > > - goto out;
> > > old_run_ctx = bpf_set_run_ctx(&run_ctx.run_ctx);
> > > item = &array->items[0];
> > > while ((prog = READ_ONCE(item->prog))) {
> > > @@ -2229,7 +2228,6 @@ bpf_prog_run_array_uprobe(const struct
> > > bpf_prog_array __rcu *array_rcu,
> > > bpf_reset_run_ctx(old_run_ctx);
> > > out:
> > > migrate_enable();
> > > - rcu_read_unlock_trace();
> > > return ret;
> > > }
> > >
> > > diff --git a/kernel/trace/trace_uprobe.c b/kernel/trace/trace_uprobe.c
> > > index fed382b7881b..87a2b8fefa90 100644
> > > --- a/kernel/trace/trace_uprobe.c
> > > +++ b/kernel/trace/trace_uprobe.c
> > > @@ -1404,7 +1404,9 @@ static void __uprobe_perf_func(struct trace_uprobe *tu,
> > > if (bpf_prog_array_valid(call)) {
> > > u32 ret;
> > >
> > > + rcu_read_lock_trace();
> > > ret = bpf_prog_run_array_uprobe(call->prog_array,
> > > regs, bpf_prog_run);
> >
> > But then this should be something like this (possibly split across
> > multiple lines with a helper variable or such):
> >
> > ret = bpf_prog_run_array_uprobe(rcu_dereference_check(call->prog_array,
> > rcu_read_lock_trace_held()), regs, bpf_prog_run);
>
> Yeah, absolutely, forgot to move the RCU dereference part, good catch!
> But I wouldn't do the _check() variant here, literally the previous
> line does rcu_read_trace_lock(), so this check part seems like just
> unnecessary verboseness, I'd go with a simple rcu_dereference().
rcu_dereference() is not legal there - that asserts that we are in a
normal RCU read-side critical section, which we are not.
rcu_dereference_raw() would be, but I think it is nice to document the
semantics to make it explicit under which lock we're operating.
I'll send a v3 in a bit after testing it.
